1 /* Copyright (c) 2013 Coraid, Inc. See COPYING for GPL terms. */ 2 /* 3 * aoecmd.c 4 * Filesystem request handling methods 5 */ 6 7 #include <linux/ata.h> 8 #include <linux/slab.h> 9 #include <linux/hdreg.h> 10 #include <linux/blkdev.h> 11 #include <linux/skbuff.h> 12 #include <linux/netdevice.h> 13 #include <linux/genhd.h> 14 #include <linux/moduleparam.h> 15 #include <linux/workqueue.h> 16 #include <linux/kthread.h> 17 #include <net/net_namespace.h> 18 #include <asm/unaligned.h> 19 #include <linux/uio.h> 20 #include "aoe.h" 21 22 #define MAXIOC (8192) /* default meant to avoid most soft lockups */ 23 24 static void ktcomplete(struct frame *, struct sk_buff *); 25 static int count_targets(struct aoedev *d, int *untainted); 26 27 static struct buf *nextbuf(struct aoedev *); 28 29 static int aoe_deadsecs = 60 * 3; 30 module_param(aoe_deadsecs, int, 0644); 31 MODULE_PARM_DESC(aoe_deadsecs, "After aoe_deadsecs seconds, give up and fail dev."); 32 33 static int aoe_maxout = 64; 34 module_param(aoe_maxout, int, 0644); 35 MODULE_PARM_DESC(aoe_maxout, 36 "Only aoe_maxout outstanding packets for every MAC on eX.Y."); 37 38 /* The number of online cpus during module initialization gives us a 39 * convenient heuristic cap on the parallelism used for ktio threads 40 * doing I/O completion. It is not important that the cap equal the 41 * actual number of running CPUs at any given time, but because of CPU 42 * hotplug, we take care to use ncpus instead of using 43 * num_online_cpus() after module initialization. 44 */ 45 static int ncpus; 46 47 /* mutex lock used for synchronization while thread spawning */ 48 static DEFINE_MUTEX(ktio_spawn_lock); 49 50 static wait_queue_head_t *ktiowq; 51 static struct ktstate *kts; 52 53 /* io completion queue */ 54 struct iocq_ktio { 55 struct list_head head; 56 spinlock_t lock; 57 }; 58 static struct iocq_ktio *iocq; 59 60 static struct page *empty_page; 61 62 static struct sk_buff * 63 new_skb(ulong len) 64 { 65 struct sk_buff *skb; 66 67 skb = alloc_skb(len + MAX_HEADER, GFP_ATOMIC); 68 if (skb) { 69 skb_reserve(skb, MAX_HEADER); 70 skb_reset_mac_header(skb); 71 skb_reset_network_header(skb); 72 skb->protocol = __constant_htons(ETH_P_AOE); 73 skb_checksum_none_assert(skb); 74 } 75 return skb; 76 } 77 78 static struct frame * 79 getframe_deferred(struct aoedev *d, u32 tag) 80 { 81 struct list_head *head, *pos, *nx; 82 struct frame *f; 83 84 head = &d->rexmitq; 85 list_for_each_safe(pos, nx, head) { 86 f = list_entry(pos, struct frame, head); 87 if (f->tag == tag) { 88 list_del(pos); 89 return f; 90 } 91 } 92 return NULL; 93 } 94 95 static struct frame * 96 getframe(struct aoedev *d, u32 tag) 97 { 98 struct frame *f; 99 struct list_head *head, *pos, *nx; 100 u32 n; 101 102 n = tag % NFACTIVE; 103 head = &d->factive[n]; 104 list_for_each_safe(pos, nx, head) { 105 f = list_entry(pos, struct frame, head); 106 if (f->tag == tag) { 107 list_del(pos); 108 return f; 109 } 110 } 111 return NULL; 112 } 113 114 /* 115 * Leave the top bit clear so we have tagspace for userland. 116 * The bottom 16 bits are the xmit tick for rexmit/rttavg processing. 117 * This driver reserves tag -1 to mean "unused frame." 118 */ 119 static int 120 newtag(struct aoedev *d) 121 { 122 register ulong n; 123 124 n = jiffies & 0xffff; 125 return n |= (++d->lasttag & 0x7fff) << 16; 126 } 127 128 static u32 129 aoehdr_atainit(struct aoedev *d, struct aoetgt *t, struct aoe_hdr *h) 130 { 131 u32 host_tag = newtag(d); 132 133 memcpy(h->src, t->ifp->nd->dev_addr, sizeof h->src); 134 memcpy(h->dst, t->addr, sizeof h->dst); 135 h->type = __constant_cpu_to_be16(ETH_P_AOE); 136 h->verfl = AOE_HVER; 137 h->major = cpu_to_be16(d->aoemajor); 138 h->minor = d->aoeminor; 139 h->cmd = AOECMD_ATA; 140 h->tag = cpu_to_be32(host_tag); 141 142 return host_tag; 143 } 144 145 static inline void 146 put_lba(struct aoe_atahdr *ah, sector_t lba) 147 { 148 ah->lba0 = lba; 149 ah->lba1 = lba >>= 8; 150 ah->lba2 = lba >>= 8; 151 ah->lba3 = lba >>= 8; 152 ah->lba4 = lba >>= 8; 153 ah->lba5 = lba >>= 8; 154 } 155 156 static struct aoeif * 157 ifrotate(struct aoetgt *t) 158 { 159 struct aoeif *ifp; 160 161 ifp = t->ifp; 162 ifp++; 163 if (ifp >= &t->ifs[NAOEIFS] || ifp->nd == NULL) 164 ifp = t->ifs; 165 if (ifp->nd == NULL) 166 return NULL; 167 return t->ifp = ifp; 168 } 169 170 static void 171 skb_pool_put(struct aoedev *d, struct sk_buff *skb) 172 { 173 __skb_queue_tail(&d->skbpool, skb); 174 } 175 176 static struct sk_buff * 177 skb_pool_get(struct aoedev *d) 178 { 179 struct sk_buff *skb = skb_peek(&d->skbpool); 180 181 if (skb && atomic_read(&skb_shinfo(skb)->dataref) == 1) { 182 __skb_unlink(skb, &d->skbpool); 183 return skb; 184 } 185 if (skb_queue_len(&d->skbpool) < NSKBPOOLMAX && 186 (skb = new_skb(ETH_ZLEN))) 187 return skb; 188 189 return NULL; 190 } 191 192 void 193 aoe_freetframe(struct frame *f) 194 { 195 struct aoetgt *t; 196 197 t = f->t; 198 f->buf = NULL; 199 f->lba = 0; 200 f->bv = NULL; 201 f->r_skb = NULL; 202 f->flags = 0; 203 list_add(&f->head, &t->ffree); 204 } 205 206 static struct frame * 207 newtframe(struct aoedev *d, struct aoetgt *t) 208 { 209 struct frame *f; 210 struct sk_buff *skb; 211 struct list_head *pos; 212 213 if (list_empty(&t->ffree)) { 214 if (t->falloc >= NSKBPOOLMAX*2) 215 return NULL; 216 f = kcalloc(1, sizeof(*f), GFP_ATOMIC); 217 if (f == NULL) 218 return NULL; 219 t->falloc++; 220 f->t = t; 221 } else { 222 pos = t->ffree.next; 223 list_del(pos); 224 f = list_entry(pos, struct frame, head); 225 } 226 227 skb = f->skb; 228 if (skb == NULL) { 229 f->skb = skb = new_skb(ETH_ZLEN); 230 if (!skb) { 231 bail: aoe_freetframe(f); 232 return NULL; 233 } 234 } 235 236 if (atomic_read(&skb_shinfo(skb)->dataref) != 1) { 237 skb = skb_pool_get(d); 238 if (skb == NULL) 239 goto bail; 240 skb_pool_put(d, f->skb); 241 f->skb = skb; 242 } 243 244 skb->truesize -= skb->data_len; 245 skb_shinfo(skb)->nr_frags = skb->data_len = 0; 246 skb_trim(skb, 0); 247 return f; 248 } 249 250 static struct frame * 251 newframe(struct aoedev *d) 252 { 253 struct frame *f; 254 struct aoetgt *t, **tt; 255 int totout = 0; 256 int use_tainted; 257 int has_untainted; 258 259 if (!d->targets || !d->targets[0]) { 260 printk(KERN_ERR "aoe: NULL TARGETS!\n"); 261 return NULL; 262 } 263 tt = d->tgt; /* last used target */ 264 for (use_tainted = 0, has_untainted = 0;;) { 265 tt++; 266 if (tt >= &d->targets[d->ntargets] || !*tt) 267 tt = d->targets; 268 t = *tt; 269 if (!t->taint) { 270 has_untainted = 1; 271 totout += t->nout; 272 } 273 if (t->nout < t->maxout 274 && (use_tainted || !t->taint) 275 && t->ifp->nd) { 276 f = newtframe(d, t); 277 if (f) { 278 ifrotate(t); 279 d->tgt = tt; 280 return f; 281 } 282 } 283 if (tt == d->tgt) { /* we've looped and found nada */ 284 if (!use_tainted && !has_untainted) 285 use_tainted = 1; 286 else 287 break; 288 } 289 } 290 if (totout == 0) { 291 d->kicked++; 292 d->flags |= DEVFL_KICKME; 293 } 294 return NULL; 295 } 296 297 static void 298 skb_fillup(struct sk_buff *skb, struct bio_vec *bv, ulong off, ulong cnt) 299 { 300 int frag = 0; 301 ulong fcnt; 302 loop: 303 fcnt = bv->bv_len - (off - bv->bv_offset); 304 if (fcnt > cnt) 305 fcnt = cnt; 306 skb_fill_page_desc(skb, frag++, bv->bv_page, off, fcnt); 307 cnt -= fcnt; 308 if (cnt <= 0) 309 return; 310 bv++; 311 off = bv->bv_offset; 312 goto loop; 313 } 314 315 static void 316 fhash(struct frame *f) 317 { 318 struct aoedev *d = f->t->d; 319 u32 n; 320 321 n = f->tag % NFACTIVE; 322 list_add_tail(&f->head, &d->factive[n]); 323 } 324 325 static void 326 ata_rw_frameinit(struct frame *f) 327 { 328 struct aoetgt *t; 329 struct aoe_hdr *h; 330 struct aoe_atahdr *ah; 331 struct sk_buff *skb; 332 char writebit, extbit; 333 334 skb = f->skb; 335 h = (struct aoe_hdr *) skb_mac_header(skb); 336 ah = (struct aoe_atahdr *) (h + 1); 337 skb_put(skb, sizeof(*h) + sizeof(*ah)); 338 memset(h, 0, skb->len); 339 340 writebit = 0x10; 341 extbit = 0x4; 342 343 t = f->t; 344 f->tag = aoehdr_atainit(t->d, t, h); 345 fhash(f); 346 t->nout++; 347 f->waited = 0; 348 f->waited_total = 0; 349 if (f->buf) 350 f->lba = f->buf->sector; 351 352 /* set up ata header */ 353 ah->scnt = f->bcnt >> 9; 354 put_lba(ah, f->lba); 355 if (t->d->flags & DEVFL_EXT) { 356 ah->aflags |= AOEAFL_EXT; 357 } else { 358 extbit = 0; 359 ah->lba3 &= 0x0f; 360 ah->lba3 |= 0xe0; /* LBA bit + obsolete 0xa0 */ 361 } 362 if (f->buf && bio_data_dir(f->buf->bio) == WRITE) { 363 skb_fillup(skb, f->bv, f->bv_off, f->bcnt); 364 ah->aflags |= AOEAFL_WRITE; 365 skb->len += f->bcnt; 366 skb->data_len = f->bcnt; 367 skb->truesize += f->bcnt; 368 t->wpkts++; 369 } else { 370 t->rpkts++; 371 writebit = 0; 372 } 373 374 ah->cmdstat = ATA_CMD_PIO_READ | writebit | extbit; 375 skb->dev = t->ifp->nd; 376 } 377 378 static int 379 aoecmd_ata_rw(struct aoedev *d) 380 { 381 struct frame *f; 382 struct buf *buf; 383 struct aoetgt *t; 384 struct sk_buff *skb; 385 struct sk_buff_head queue; 386 ulong bcnt, fbcnt; 387 388 buf = nextbuf(d); 389 if (buf == NULL) 390 return 0; 391 f = newframe(d); 392 if (f == NULL) 393 return 0; 394 t = *d->tgt; 395 bcnt = d->maxbcnt; 396 if (bcnt == 0) 397 bcnt = DEFAULTBCNT; 398 if (bcnt > buf->resid) 399 bcnt = buf->resid; 400 fbcnt = bcnt; 401 f->bv = buf->bv; 402 f->bv_off = f->bv->bv_offset + (f->bv->bv_len - buf->bv_resid); 403 do { 404 if (fbcnt < buf->bv_resid) { 405 buf->bv_resid -= fbcnt; 406 buf->resid -= fbcnt; 407 break; 408 } 409 fbcnt -= buf->bv_resid; 410 buf->resid -= buf->bv_resid; 411 if (buf->resid == 0) { 412 d->ip.buf = NULL; 413 break; 414 } 415 buf->bv++; 416 buf->bv_resid = buf->bv->bv_len; 417 WARN_ON(buf->bv_resid == 0); 418 } while (fbcnt); 419 420 /* initialize the headers & frame */ 421 f->buf = buf; 422 f->bcnt = bcnt; 423 ata_rw_frameinit(f); 424 425 /* mark all tracking fields and load out */ 426 buf->nframesout += 1; 427 buf->sector += bcnt >> 9; 428 429 skb = skb_clone(f->skb, GFP_ATOMIC); 430 if (skb) { 431 do_gettimeofday(&f->sent); 432 f->sent_jiffs = (u32) jiffies; 433 __skb_queue_head_init(&queue); 434 __skb_queue_tail(&queue, skb); 435 aoenet_xmit(&queue); 436 } 437 return 1; 438 } 439 440 /* some callers cannot sleep, and they can call this function, 441 * transmitting the packets later, when interrupts are on 442 */ 443 static void 444 aoecmd_cfg_pkts(ushort aoemajor, unsigned char aoeminor, struct sk_buff_head *queue) 445 { 446 struct aoe_hdr *h; 447 struct aoe_cfghdr *ch; 448 struct sk_buff *skb; 449 struct net_device *ifp; 450 451 rcu_read_lock(); 452 for_each_netdev_rcu(&init_net, ifp) { 453 dev_hold(ifp); 454 if (!is_aoe_netif(ifp)) 455 goto cont; 456 457 skb = new_skb(sizeof *h + sizeof *ch); 458 if (skb == NULL) { 459 printk(KERN_INFO "aoe: skb alloc failure\n"); 460 goto cont; 461 } 462 skb_put(skb, sizeof *h + sizeof *ch); 463 skb->dev = ifp; 464 __skb_queue_tail(queue, skb); 465 h = (struct aoe_hdr *) skb_mac_header(skb); 466 memset(h, 0, sizeof *h + sizeof *ch); 467 468 memset(h->dst, 0xff, sizeof h->dst); 469 memcpy(h->src, ifp->dev_addr, sizeof h->src); 470 h->type = __constant_cpu_to_be16(ETH_P_AOE); 471 h->verfl = AOE_HVER; 472 h->major = cpu_to_be16(aoemajor); 473 h->minor = aoeminor; 474 h->cmd = AOECMD_CFG; 475 476 cont: 477 dev_put(ifp); 478 } 479 rcu_read_unlock(); 480 } 481 482 static void 483 resend(struct aoedev *d, struct frame *f) 484 { 485 struct sk_buff *skb; 486 struct sk_buff_head queue; 487 struct aoe_hdr *h; 488 struct aoe_atahdr *ah; 489 struct aoetgt *t; 490 char buf[128]; 491 u32 n; 492 493 t = f->t; 494 n = newtag(d); 495 skb = f->skb; 496 if (ifrotate(t) == NULL) { 497 /* probably can't happen, but set it up to fail anyway */ 498 pr_info("aoe: resend: no interfaces to rotate to.\n"); 499 ktcomplete(f, NULL); 500 return; 501 } 502 h = (struct aoe_hdr *) skb_mac_header(skb); 503 ah = (struct aoe_atahdr *) (h+1); 504 505 if (!(f->flags & FFL_PROBE)) { 506 snprintf(buf, sizeof(buf), 507 "%15s e%ld.%d oldtag=%08x@%08lx newtag=%08x s=%pm d=%pm nout=%d\n", 508 "retransmit", d->aoemajor, d->aoeminor, 509 f->tag, jiffies, n, 510 h->src, h->dst, t->nout); 511 aoechr_error(buf); 512 } 513 514 f->tag = n; 515 fhash(f); 516 h->tag = cpu_to_be32(n); 517 memcpy(h->dst, t->addr, sizeof h->dst); 518 memcpy(h->src, t->ifp->nd->dev_addr, sizeof h->src); 519 520 skb->dev = t->ifp->nd; 521 skb = skb_clone(skb, GFP_ATOMIC); 522 if (skb == NULL) 523 return; 524 do_gettimeofday(&f->sent); 525 f->sent_jiffs = (u32) jiffies; 526 __skb_queue_head_init(&queue); 527 __skb_queue_tail(&queue, skb); 528 aoenet_xmit(&queue); 529 } 530 531 static int 532 tsince_hr(struct frame *f) 533 { 534 struct timeval now; 535 int n; 536 537 do_gettimeofday(&now); 538 n = now.tv_usec - f->sent.tv_usec; 539 n += (now.tv_sec - f->sent.tv_sec) * USEC_PER_SEC; 540 541 if (n < 0) 542 n = -n; 543 544 /* For relatively long periods, use jiffies to avoid 545 * discrepancies caused by updates to the system time. 546 * 547 * On system with HZ of 1000, 32-bits is over 49 days 548 * worth of jiffies, or over 71 minutes worth of usecs. 549 * 550 * Jiffies overflow is handled by subtraction of unsigned ints: 551 * (gdb) print (unsigned) 2 - (unsigned) 0xfffffffe 552 * $3 = 4 553 * (gdb) 554 */ 555 if (n > USEC_PER_SEC / 4) { 556 n = ((u32) jiffies) - f->sent_jiffs; 557 n *= USEC_PER_SEC / HZ; 558 } 559 560 return n; 561 } 562 563 static int 564 tsince(u32 tag) 565 { 566 int n; 567 568 n = jiffies & 0xffff; 569 n -= tag & 0xffff; 570 if (n < 0) 571 n += 1<<16; 572 return jiffies_to_usecs(n + 1); 573 } 574 575 static struct aoeif * 576 getif(struct aoetgt *t, struct net_device *nd) 577 { 578 struct aoeif *p, *e; 579 580 p = t->ifs; 581 e = p + NAOEIFS; 582 for (; p < e; p++) 583 if (p->nd == nd) 584 return p; 585 return NULL; 586 } 587 588 static void 589 ejectif(struct aoetgt *t, struct aoeif *ifp) 590 { 591 struct aoeif *e; 592 struct net_device *nd; 593 ulong n; 594 595 nd = ifp->nd; 596 e = t->ifs + NAOEIFS - 1; 597 n = (e - ifp) * sizeof *ifp; 598 memmove(ifp, ifp+1, n); 599 e->nd = NULL; 600 dev_put(nd); 601 } 602 603 static struct frame * 604 reassign_frame(struct frame *f) 605 { 606 struct frame *nf; 607 struct sk_buff *skb; 608 609 nf = newframe(f->t->d); 610 if (!nf) 611 return NULL; 612 if (nf->t == f->t) { 613 aoe_freetframe(nf); 614 return NULL; 615 } 616 617 skb = nf->skb; 618 nf->skb = f->skb; 619 nf->buf = f->buf; 620 nf->bcnt = f->bcnt; 621 nf->lba = f->lba; 622 nf->bv = f->bv; 623 nf->bv_off = f->bv_off; 624 nf->waited = 0; 625 nf->waited_total = f->waited_total; 626 nf->sent = f->sent; 627 nf->sent_jiffs = f->sent_jiffs; 628 f->skb = skb; 629 630 return nf; 631 } 632 633 static void 634 probe(struct aoetgt *t) 635 { 636 struct aoedev *d; 637 struct frame *f; 638 struct sk_buff *skb; 639 struct sk_buff_head queue; 640 size_t n, m; 641 int frag; 642 643 d = t->d; 644 f = newtframe(d, t); 645 if (!f) { 646 pr_err("%s %pm for e%ld.%d: %s\n", 647 "aoe: cannot probe remote address", 648 t->addr, 649 (long) d->aoemajor, d->aoeminor, 650 "no frame available"); 651 return; 652 } 653 f->flags |= FFL_PROBE; 654 ifrotate(t); 655 f->bcnt = t->d->maxbcnt ? t->d->maxbcnt : DEFAULTBCNT; 656 ata_rw_frameinit(f); 657 skb = f->skb; 658 for (frag = 0, n = f->bcnt; n > 0; ++frag, n -= m) { 659 if (n < PAGE_SIZE) 660 m = n; 661 else 662 m = PAGE_SIZE; 663 skb_fill_page_desc(skb, frag, empty_page, 0, m); 664 } 665 skb->len += f->bcnt; 666 skb->data_len = f->bcnt; 667 skb->truesize += f->bcnt; 668 669 skb = skb_clone(f->skb, GFP_ATOMIC); 670 if (skb) { 671 do_gettimeofday(&f->sent); 672 f->sent_jiffs = (u32) jiffies; 673 __skb_queue_head_init(&queue); 674 __skb_queue_tail(&queue, skb); 675 aoenet_xmit(&queue); 676 } 677 } 678 679 static long 680 rto(struct aoedev *d) 681 { 682 long t; 683 684 t = 2 * d->rttavg >> RTTSCALE; 685 t += 8 * d->rttdev >> RTTDSCALE; 686 if (t == 0) 687 t = 1; 688 689 return t; 690 } 691 692 static void 693 rexmit_deferred(struct aoedev *d) 694 { 695 struct aoetgt *t; 696 struct frame *f; 697 struct frame *nf; 698 struct list_head *pos, *nx, *head; 699 int since; 700 int untainted; 701 702 count_targets(d, &untainted); 703 704 head = &d->rexmitq; 705 list_for_each_safe(pos, nx, head) { 706 f = list_entry(pos, struct frame, head); 707 t = f->t; 708 if (t->taint) { 709 if (!(f->flags & FFL_PROBE)) { 710 nf = reassign_frame(f); 711 if (nf) { 712 if (t->nout_probes == 0 713 && untainted > 0) { 714 probe(t); 715 t->nout_probes++; 716 } 717 list_replace(&f->head, &nf->head); 718 pos = &nf->head; 719 aoe_freetframe(f); 720 f = nf; 721 t = f->t; 722 } 723 } else if (untainted < 1) { 724 /* don't probe w/o other untainted aoetgts */ 725 goto stop_probe; 726 } else if (tsince_hr(f) < t->taint * rto(d)) { 727 /* reprobe slowly when taint is high */ 728 continue; 729 } 730 } else if (f->flags & FFL_PROBE) { 731 stop_probe: /* don't probe untainted aoetgts */ 732 list_del(pos); 733 aoe_freetframe(f); 734 /* leaving d->kicked, because this is routine */ 735 f->t->d->flags |= DEVFL_KICKME; 736 continue; 737 } 738 if (t->nout >= t->maxout) 739 continue; 740 list_del(pos); 741 t->nout++; 742 if (f->flags & FFL_PROBE) 743 t->nout_probes++; 744 since = tsince_hr(f); 745 f->waited += since; 746 f->waited_total += since; 747 resend(d, f); 748 } 749 } 750 751 /* An aoetgt accumulates demerits quickly, and successful 752 * probing redeems the aoetgt slowly. 753 */ 754 static void 755 scorn(struct aoetgt *t) 756 { 757 int n; 758 759 n = t->taint++; 760 t->taint += t->taint * 2; 761 if (n > t->taint) 762 t->taint = n; 763 if (t->taint > MAX_TAINT) 764 t->taint = MAX_TAINT; 765 } 766 767 static int 768 count_targets(struct aoedev *d, int *untainted) 769 { 770 int i, good; 771 772 for (i = good = 0; i < d->ntargets && d->targets[i]; ++i) 773 if (d->targets[i]->taint == 0) 774 good++; 775 776 if (untainted) 777 *untainted = good; 778 return i; 779 } 780 781 static void 782 rexmit_timer(ulong vp) 783 { 784 struct aoedev *d; 785 struct aoetgt *t; 786 struct aoeif *ifp; 787 struct frame *f; 788 struct list_head *head, *pos, *nx; 789 LIST_HEAD(flist); 790 register long timeout; 791 ulong flags, n; 792 int i; 793 int utgts; /* number of aoetgt descriptors (not slots) */ 794 int since; 795 796 d = (struct aoedev *) vp; 797 798 spin_lock_irqsave(&d->lock, flags); 799 800 /* timeout based on observed timings and variations */ 801 timeout = rto(d); 802 803 utgts = count_targets(d, NULL); 804 805 if (d->flags & DEVFL_TKILL) { 806 spin_unlock_irqrestore(&d->lock, flags); 807 return; 808 } 809 810 /* collect all frames to rexmit into flist */ 811 for (i = 0; i < NFACTIVE; i++) { 812 head = &d->factive[i]; 813 list_for_each_safe(pos, nx, head) { 814 f = list_entry(pos, struct frame, head); 815 if (tsince_hr(f) < timeout) 816 break; /* end of expired frames */ 817 /* move to flist for later processing */ 818 list_move_tail(pos, &flist); 819 } 820 } 821 822 /* process expired frames */ 823 while (!list_empty(&flist)) { 824 pos = flist.next; 825 f = list_entry(pos, struct frame, head); 826 since = tsince_hr(f); 827 n = f->waited_total + since; 828 n /= USEC_PER_SEC; 829 if (aoe_deadsecs 830 && n > aoe_deadsecs 831 && !(f->flags & FFL_PROBE)) { 832 /* Waited too long. Device failure. 833 * Hang all frames on first hash bucket for downdev 834 * to clean up. 835 */ 836 list_splice(&flist, &d->factive[0]); 837 aoedev_downdev(d); 838 goto out; 839 } 840 841 t = f->t; 842 n = f->waited + since; 843 n /= USEC_PER_SEC; 844 if (aoe_deadsecs && utgts > 0 845 && (n > aoe_deadsecs / utgts || n > HARD_SCORN_SECS)) 846 scorn(t); /* avoid this target */ 847 848 if (t->maxout != 1) { 849 t->ssthresh = t->maxout / 2; 850 t->maxout = 1; 851 } 852 853 if (f->flags & FFL_PROBE) { 854 t->nout_probes--; 855 } else { 856 ifp = getif(t, f->skb->dev); 857 if (ifp && ++ifp->lost > (t->nframes << 1) 858 && (ifp != t->ifs || t->ifs[1].nd)) { 859 ejectif(t, ifp); 860 ifp = NULL; 861 } 862 } 863 list_move_tail(pos, &d->rexmitq); 864 t->nout--; 865 } 866 rexmit_deferred(d); 867 868 out: 869 if ((d->flags & DEVFL_KICKME) && d->blkq) { 870 d->flags &= ~DEVFL_KICKME; 871 d->blkq->request_fn(d->blkq); 872 } 873 874 d->timer.expires = jiffies + TIMERTICK; 875 add_timer(&d->timer); 876 877 spin_unlock_irqrestore(&d->lock, flags); 878 } 879 880 static unsigned long 881 rqbiocnt(struct request *r) 882 { 883 struct bio *bio; 884 unsigned long n = 0; 885 886 __rq_for_each_bio(bio, r) 887 n++; 888 return n; 889 } 890 891 /* This can be removed if we are certain that no users of the block 892 * layer will ever use zero-count pages in bios. Otherwise we have to 893 * protect against the put_page sometimes done by the network layer. 894 * 895 * See http://oss.sgi.com/archives/xfs/2007-01/msg00594.html for 896 * discussion. 897 * 898 * We cannot use get_page in the workaround, because it insists on a 899 * positive page count as a precondition. So we use _count directly. 900 */ 901 static void 902 bio_pageinc(struct bio *bio) 903 { 904 struct bio_vec *bv; 905 struct page *page; 906 int i; 907 908 bio_for_each_segment(bv, bio, i) { 909 page = bv->bv_page; 910 /* Non-zero page count for non-head members of 911 * compound pages is no longer allowed by the kernel, 912 * but this has never been seen here. 913 */ 914 if (unlikely(PageCompound(page))) 915 if (compound_trans_head(page) != page) { 916 pr_crit("page tail used for block I/O\n"); 917 BUG(); 918 } 919 atomic_inc(&page->_count); 920 } 921 } 922 923 static void 924 bio_pagedec(struct bio *bio) 925 { 926 struct bio_vec *bv; 927 int i; 928 929 bio_for_each_segment(bv, bio, i) 930 atomic_dec(&bv->bv_page->_count); 931 } 932 933 static void 934 bufinit(struct buf *buf, struct request *rq, struct bio *bio) 935 { 936 memset(buf, 0, sizeof(*buf)); 937 buf->rq = rq; 938 buf->bio = bio; 939 buf->resid = bio->bi_size; 940 buf->sector = bio->bi_sector; 941 bio_pageinc(bio); 942 buf->bv = bio_iovec(bio); 943 buf->bv_resid = buf->bv->bv_len; 944 WARN_ON(buf->bv_resid == 0); 945 } 946 947 static struct buf * 948 nextbuf(struct aoedev *d) 949 { 950 struct request *rq; 951 struct request_queue *q; 952 struct buf *buf; 953 struct bio *bio; 954 955 q = d->blkq; 956 if (q == NULL) 957 return NULL; /* initializing */ 958 if (d->ip.buf) 959 return d->ip.buf; 960 rq = d->ip.rq; 961 if (rq == NULL) { 962 rq = blk_peek_request(q); 963 if (rq == NULL) 964 return NULL; 965 blk_start_request(rq); 966 d->ip.rq = rq; 967 d->ip.nxbio = rq->bio; 968 rq->special = (void *) rqbiocnt(rq); 969 } 970 buf = mempool_alloc(d->bufpool, GFP_ATOMIC); 971 if (buf == NULL) { 972 pr_err("aoe: nextbuf: unable to mempool_alloc!\n"); 973 return NULL; 974 } 975 bio = d->ip.nxbio; 976 bufinit(buf, rq, bio); 977 bio = bio->bi_next; 978 d->ip.nxbio = bio; 979 if (bio == NULL) 980 d->ip.rq = NULL; 981 return d->ip.buf = buf; 982 } 983 984 /* enters with d->lock held */ 985 void 986 aoecmd_work(struct aoedev *d) 987 { 988 rexmit_deferred(d); 989 while (aoecmd_ata_rw(d)) 990 ; 991 } 992 993 /* this function performs work that has been deferred until sleeping is OK 994 */ 995 void 996 aoecmd_sleepwork(struct work_struct *work) 997 { 998 struct aoedev *d = container_of(work, struct aoedev, work); 999 struct block_device *bd; 1000 u64 ssize; 1001 1002 if (d->flags & DEVFL_GDALLOC) 1003 aoeblk_gdalloc(d); 1004 1005 if (d->flags & DEVFL_NEWSIZE) { 1006 ssize = get_capacity(d->gd); 1007 bd = bdget_disk(d->gd, 0); 1008 if (bd) { 1009 mutex_lock(&bd->bd_inode->i_mutex); 1010 i_size_write(bd->bd_inode, (loff_t)ssize<<9); 1011 mutex_unlock(&bd->bd_inode->i_mutex); 1012 bdput(bd); 1013 } 1014 spin_lock_irq(&d->lock); 1015 d->flags |= DEVFL_UP; 1016 d->flags &= ~DEVFL_NEWSIZE; 1017 spin_unlock_irq(&d->lock); 1018 } 1019 } 1020 1021 static void 1022 ata_ident_fixstring(u16 *id, int ns) 1023 { 1024 u16 s; 1025 1026 while (ns-- > 0) { 1027 s = *id; 1028 *id++ = s >> 8 | s << 8; 1029 } 1030 } 1031 1032 static void 1033 ataid_complete(struct aoedev *d, struct aoetgt *t, unsigned char *id) 1034 { 1035 u64 ssize; 1036 u16 n; 1037 1038 /* word 83: command set supported */ 1039 n = get_unaligned_le16(&id[83 << 1]); 1040 1041 /* word 86: command set/feature enabled */ 1042 n |= get_unaligned_le16(&id[86 << 1]); 1043 1044 if (n & (1<<10)) { /* bit 10: LBA 48 */ 1045 d->flags |= DEVFL_EXT; 1046 1047 /* word 100: number lba48 sectors */ 1048 ssize = get_unaligned_le64(&id[100 << 1]); 1049 1050 /* set as in ide-disk.c:init_idedisk_capacity */ 1051 d->geo.cylinders = ssize; 1052 d->geo.cylinders /= (255 * 63); 1053 d->geo.heads = 255; 1054 d->geo.sectors = 63; 1055 } else { 1056 d->flags &= ~DEVFL_EXT; 1057 1058 /* number lba28 sectors */ 1059 ssize = get_unaligned_le32(&id[60 << 1]); 1060 1061 /* NOTE: obsolete in ATA 6 */ 1062 d->geo.cylinders = get_unaligned_le16(&id[54 << 1]); 1063 d->geo.heads = get_unaligned_le16(&id[55 << 1]); 1064 d->geo.sectors = get_unaligned_le16(&id[56 << 1]); 1065 } 1066 1067 ata_ident_fixstring((u16 *) &id[10<<1], 10); /* serial */ 1068 ata_ident_fixstring((u16 *) &id[23<<1], 4); /* firmware */ 1069 ata_ident_fixstring((u16 *) &id[27<<1], 20); /* model */ 1070 memcpy(d->ident, id, sizeof(d->ident)); 1071 1072 if (d->ssize != ssize) 1073 printk(KERN_INFO 1074 "aoe: %pm e%ld.%d v%04x has %llu sectors\n", 1075 t->addr, 1076 d->aoemajor, d->aoeminor, 1077 d->fw_ver, (long long)ssize); 1078 d->ssize = ssize; 1079 d->geo.start = 0; 1080 if (d->flags & (DEVFL_GDALLOC|DEVFL_NEWSIZE)) 1081 return; 1082 if (d->gd != NULL) { 1083 set_capacity(d->gd, ssize); 1084 d->flags |= DEVFL_NEWSIZE; 1085 } else 1086 d->flags |= DEVFL_GDALLOC; 1087 schedule_work(&d->work); 1088 } 1089 1090 static void 1091 calc_rttavg(struct aoedev *d, struct aoetgt *t, int rtt) 1092 { 1093 register long n; 1094 1095 n = rtt; 1096 1097 /* cf. Congestion Avoidance and Control, Jacobson & Karels, 1988 */ 1098 n -= d->rttavg >> RTTSCALE; 1099 d->rttavg += n; 1100 if (n < 0) 1101 n = -n; 1102 n -= d->rttdev >> RTTDSCALE; 1103 d->rttdev += n; 1104 1105 if (!t || t->maxout >= t->nframes) 1106 return; 1107 if (t->maxout < t->ssthresh) 1108 t->maxout += 1; 1109 else if (t->nout == t->maxout && t->next_cwnd-- == 0) { 1110 t->maxout += 1; 1111 t->next_cwnd = t->maxout; 1112 } 1113 } 1114 1115 static struct aoetgt * 1116 gettgt(struct aoedev *d, char *addr) 1117 { 1118 struct aoetgt **t, **e; 1119 1120 t = d->targets; 1121 e = t + d->ntargets; 1122 for (; t < e && *t; t++) 1123 if (memcmp((*t)->addr, addr, sizeof((*t)->addr)) == 0) 1124 return *t; 1125 return NULL; 1126 } 1127 1128 static void 1129 bvcpy(struct bio_vec *bv, ulong off, struct sk_buff *skb, long cnt) 1130 { 1131 ulong fcnt; 1132 char *p; 1133 int soff = 0; 1134 loop: 1135 fcnt = bv->bv_len - (off - bv->bv_offset); 1136 if (fcnt > cnt) 1137 fcnt = cnt; 1138 p = page_address(bv->bv_page) + off; 1139 skb_copy_bits(skb, soff, p, fcnt); 1140 soff += fcnt; 1141 cnt -= fcnt; 1142 if (cnt <= 0) 1143 return; 1144 bv++; 1145 off = bv->bv_offset; 1146 goto loop; 1147 } 1148 1149 void 1150 aoe_end_request(struct aoedev *d, struct request *rq, int fastfail) 1151 { 1152 struct bio *bio; 1153 int bok; 1154 struct request_queue *q; 1155 1156 q = d->blkq; 1157 if (rq == d->ip.rq) 1158 d->ip.rq = NULL; 1159 do { 1160 bio = rq->bio; 1161 bok = !fastfail && test_bit(BIO_UPTODATE, &bio->bi_flags); 1162 } while (__blk_end_request(rq, bok ? 0 : -EIO, bio->bi_size)); 1163 1164 /* cf. http://lkml.org/lkml/2006/10/31/28 */ 1165 if (!fastfail) 1166 __blk_run_queue(q); 1167 } 1168 1169 static void 1170 aoe_end_buf(struct aoedev *d, struct buf *buf) 1171 { 1172 struct request *rq; 1173 unsigned long n; 1174 1175 if (buf == d->ip.buf) 1176 d->ip.buf = NULL; 1177 rq = buf->rq; 1178 bio_pagedec(buf->bio); 1179 mempool_free(buf, d->bufpool); 1180 n = (unsigned long) rq->special; 1181 rq->special = (void *) --n; 1182 if (n == 0) 1183 aoe_end_request(d, rq, 0); 1184 } 1185 1186 static void 1187 ktiocomplete(struct frame *f) 1188 { 1189 struct aoe_hdr *hin, *hout; 1190 struct aoe_atahdr *ahin, *ahout; 1191 struct buf *buf; 1192 struct sk_buff *skb; 1193 struct aoetgt *t; 1194 struct aoeif *ifp; 1195 struct aoedev *d; 1196 long n; 1197 int untainted; 1198 1199 if (f == NULL) 1200 return; 1201 1202 t = f->t; 1203 d = t->d; 1204 skb = f->r_skb; 1205 buf = f->buf; 1206 if (f->flags & FFL_PROBE) 1207 goto out; 1208 if (!skb) /* just fail the buf. */ 1209 goto noskb; 1210 1211 hout = (struct aoe_hdr *) skb_mac_header(f->skb); 1212 ahout = (struct aoe_atahdr *) (hout+1); 1213 1214 hin = (struct aoe_hdr *) skb->data; 1215 skb_pull(skb, sizeof(*hin)); 1216 ahin = (struct aoe_atahdr *) skb->data; 1217 skb_pull(skb, sizeof(*ahin)); 1218 if (ahin->cmdstat & 0xa9) { /* these bits cleared on success */ 1219 pr_err("aoe: ata error cmd=%2.2Xh stat=%2.2Xh from e%ld.%d\n", 1220 ahout->cmdstat, ahin->cmdstat, 1221 d->aoemajor, d->aoeminor); 1222 noskb: if (buf) 1223 clear_bit(BIO_UPTODATE, &buf->bio->bi_flags); 1224 goto out; 1225 } 1226 1227 n = ahout->scnt << 9; 1228 switch (ahout->cmdstat) { 1229 case ATA_CMD_PIO_READ: 1230 case ATA_CMD_PIO_READ_EXT: 1231 if (skb->len < n) { 1232 pr_err("%s e%ld.%d. skb->len=%d need=%ld\n", 1233 "aoe: runt data size in read from", 1234 (long) d->aoemajor, d->aoeminor, 1235 skb->len, n); 1236 clear_bit(BIO_UPTODATE, &buf->bio->bi_flags); 1237 break; 1238 } 1239 bvcpy(f->bv, f->bv_off, skb, n); 1240 case ATA_CMD_PIO_WRITE: 1241 case ATA_CMD_PIO_WRITE_EXT: 1242 spin_lock_irq(&d->lock); 1243 ifp = getif(t, skb->dev); 1244 if (ifp) 1245 ifp->lost = 0; 1246 spin_unlock_irq(&d->lock); 1247 break; 1248 case ATA_CMD_ID_ATA: 1249 if (skb->len < 512) { 1250 pr_info("%s e%ld.%d. skb->len=%d need=512\n", 1251 "aoe: runt data size in ataid from", 1252 (long) d->aoemajor, d->aoeminor, 1253 skb->len); 1254 break; 1255 } 1256 if (skb_linearize(skb)) 1257 break; 1258 spin_lock_irq(&d->lock); 1259 ataid_complete(d, t, skb->data); 1260 spin_unlock_irq(&d->lock); 1261 break; 1262 default: 1263 pr_info("aoe: unrecognized ata command %2.2Xh for %d.%d\n", 1264 ahout->cmdstat, 1265 be16_to_cpu(get_unaligned(&hin->major)), 1266 hin->minor); 1267 } 1268 out: 1269 spin_lock_irq(&d->lock); 1270 if (t->taint > 0 1271 && --t->taint > 0 1272 && t->nout_probes == 0) { 1273 count_targets(d, &untainted); 1274 if (untainted > 0) { 1275 probe(t); 1276 t->nout_probes++; 1277 } 1278 } 1279 1280 aoe_freetframe(f); 1281 1282 if (buf && --buf->nframesout == 0 && buf->resid == 0) 1283 aoe_end_buf(d, buf); 1284 1285 spin_unlock_irq(&d->lock); 1286 aoedev_put(d); 1287 dev_kfree_skb(skb); 1288 } 1289 1290 /* Enters with iocq.lock held. 1291 * Returns true iff responses needing processing remain. 1292 */ 1293 static int 1294 ktio(int id) 1295 { 1296 struct frame *f; 1297 struct list_head *pos; 1298 int i; 1299 int actual_id; 1300 1301 for (i = 0; ; ++i) { 1302 if (i == MAXIOC) 1303 return 1; 1304 if (list_empty(&iocq[id].head)) 1305 return 0; 1306 pos = iocq[id].head.next; 1307 list_del(pos); 1308 f = list_entry(pos, struct frame, head); 1309 spin_unlock_irq(&iocq[id].lock); 1310 ktiocomplete(f); 1311 1312 /* Figure out if extra threads are required. */ 1313 actual_id = f->t->d->aoeminor % ncpus; 1314 1315 if (!kts[actual_id].active) { 1316 BUG_ON(id != 0); 1317 mutex_lock(&ktio_spawn_lock); 1318 if (!kts[actual_id].active 1319 && aoe_ktstart(&kts[actual_id]) == 0) 1320 kts[actual_id].active = 1; 1321 mutex_unlock(&ktio_spawn_lock); 1322 } 1323 spin_lock_irq(&iocq[id].lock); 1324 } 1325 } 1326 1327 static int 1328 kthread(void *vp) 1329 { 1330 struct ktstate *k; 1331 DECLARE_WAITQUEUE(wait, current); 1332 int more; 1333 1334 k = vp; 1335 current->flags |= PF_NOFREEZE; 1336 set_user_nice(current, -10); 1337 complete(&k->rendez); /* tell spawner we're running */ 1338 do { 1339 spin_lock_irq(k->lock); 1340 more = k->fn(k->id); 1341 if (!more) { 1342 add_wait_queue(k->waitq, &wait); 1343 __set_current_state(TASK_INTERRUPTIBLE); 1344 } 1345 spin_unlock_irq(k->lock); 1346 if (!more) { 1347 schedule(); 1348 remove_wait_queue(k->waitq, &wait); 1349 } else 1350 cond_resched(); 1351 } while (!kthread_should_stop()); 1352 complete(&k->rendez); /* tell spawner we're stopping */ 1353 return 0; 1354 } 1355 1356 void 1357 aoe_ktstop(struct ktstate *k) 1358 { 1359 kthread_stop(k->task); 1360 wait_for_completion(&k->rendez); 1361 } 1362 1363 int 1364 aoe_ktstart(struct ktstate *k) 1365 { 1366 struct task_struct *task; 1367 1368 init_completion(&k->rendez); 1369 task = kthread_run(kthread, k, "%s", k->name); 1370 if (task == NULL || IS_ERR(task)) 1371 return -ENOMEM; 1372 k->task = task; 1373 wait_for_completion(&k->rendez); /* allow kthread to start */ 1374 init_completion(&k->rendez); /* for waiting for exit later */ 1375 return 0; 1376 } 1377 1378 /* pass it off to kthreads for processing */ 1379 static void 1380 ktcomplete(struct frame *f, struct sk_buff *skb) 1381 { 1382 int id; 1383 ulong flags; 1384 1385 f->r_skb = skb; 1386 id = f->t->d->aoeminor % ncpus; 1387 spin_lock_irqsave(&iocq[id].lock, flags); 1388 if (!kts[id].active) { 1389 spin_unlock_irqrestore(&iocq[id].lock, flags); 1390 /* The thread with id has not been spawned yet, 1391 * so delegate the work to the main thread and 1392 * try spawning a new thread. 1393 */ 1394 id = 0; 1395 spin_lock_irqsave(&iocq[id].lock, flags); 1396 } 1397 list_add_tail(&f->head, &iocq[id].head); 1398 spin_unlock_irqrestore(&iocq[id].lock, flags); 1399 wake_up(&ktiowq[id]); 1400 } 1401 1402 struct sk_buff * 1403 aoecmd_ata_rsp(struct sk_buff *skb) 1404 { 1405 struct aoedev *d; 1406 struct aoe_hdr *h; 1407 struct frame *f; 1408 u32 n; 1409 ulong flags; 1410 char ebuf[128]; 1411 u16 aoemajor; 1412 1413 h = (struct aoe_hdr *) skb->data; 1414 aoemajor = be16_to_cpu(get_unaligned(&h->major)); 1415 d = aoedev_by_aoeaddr(aoemajor, h->minor, 0); 1416 if (d == NULL) { 1417 snprintf(ebuf, sizeof ebuf, "aoecmd_ata_rsp: ata response " 1418 "for unknown device %d.%d\n", 1419 aoemajor, h->minor); 1420 aoechr_error(ebuf); 1421 return skb; 1422 } 1423 1424 spin_lock_irqsave(&d->lock, flags); 1425 1426 n = be32_to_cpu(get_unaligned(&h->tag)); 1427 f = getframe(d, n); 1428 if (f) { 1429 calc_rttavg(d, f->t, tsince_hr(f)); 1430 f->t->nout--; 1431 if (f->flags & FFL_PROBE) 1432 f->t->nout_probes--; 1433 } else { 1434 f = getframe_deferred(d, n); 1435 if (f) { 1436 calc_rttavg(d, NULL, tsince_hr(f)); 1437 } else { 1438 calc_rttavg(d, NULL, tsince(n)); 1439 spin_unlock_irqrestore(&d->lock, flags); 1440 aoedev_put(d); 1441 snprintf(ebuf, sizeof(ebuf), 1442 "%15s e%d.%d tag=%08x@%08lx s=%pm d=%pm\n", 1443 "unexpected rsp", 1444 get_unaligned_be16(&h->major), 1445 h->minor, 1446 get_unaligned_be32(&h->tag), 1447 jiffies, 1448 h->src, 1449 h->dst); 1450 aoechr_error(ebuf); 1451 return skb; 1452 } 1453 } 1454 aoecmd_work(d); 1455 1456 spin_unlock_irqrestore(&d->lock, flags); 1457 1458 ktcomplete(f, skb); 1459 1460 /* 1461 * Note here that we do not perform an aoedev_put, as we are 1462 * leaving this reference for the ktio to release. 1463 */ 1464 return NULL; 1465 } 1466 1467 void 1468 aoecmd_cfg(ushort aoemajor, unsigned char aoeminor) 1469 { 1470 struct sk_buff_head queue; 1471 1472 __skb_queue_head_init(&queue); 1473 aoecmd_cfg_pkts(aoemajor, aoeminor, &queue); 1474 aoenet_xmit(&queue); 1475 } 1476 1477 struct sk_buff * 1478 aoecmd_ata_id(struct aoedev *d) 1479 { 1480 struct aoe_hdr *h; 1481 struct aoe_atahdr *ah; 1482 struct frame *f; 1483 struct sk_buff *skb; 1484 struct aoetgt *t; 1485 1486 f = newframe(d); 1487 if (f == NULL) 1488 return NULL; 1489 1490 t = *d->tgt; 1491 1492 /* initialize the headers & frame */ 1493 skb = f->skb; 1494 h = (struct aoe_hdr *) skb_mac_header(skb); 1495 ah = (struct aoe_atahdr *) (h+1); 1496 skb_put(skb, sizeof *h + sizeof *ah); 1497 memset(h, 0, skb->len); 1498 f->tag = aoehdr_atainit(d, t, h); 1499 fhash(f); 1500 t->nout++; 1501 f->waited = 0; 1502 f->waited_total = 0; 1503 1504 /* set up ata header */ 1505 ah->scnt = 1; 1506 ah->cmdstat = ATA_CMD_ID_ATA; 1507 ah->lba3 = 0xa0; 1508 1509 skb->dev = t->ifp->nd; 1510 1511 d->rttavg = RTTAVG_INIT; 1512 d->rttdev = RTTDEV_INIT; 1513 d->timer.function = rexmit_timer; 1514 1515 skb = skb_clone(skb, GFP_ATOMIC); 1516 if (skb) { 1517 do_gettimeofday(&f->sent); 1518 f->sent_jiffs = (u32) jiffies; 1519 } 1520 1521 return skb; 1522 } 1523 1524 static struct aoetgt ** 1525 grow_targets(struct aoedev *d) 1526 { 1527 ulong oldn, newn; 1528 struct aoetgt **tt; 1529 1530 oldn = d->ntargets; 1531 newn = oldn * 2; 1532 tt = kcalloc(newn, sizeof(*d->targets), GFP_ATOMIC); 1533 if (!tt) 1534 return NULL; 1535 memmove(tt, d->targets, sizeof(*d->targets) * oldn); 1536 d->tgt = tt + (d->tgt - d->targets); 1537 kfree(d->targets); 1538 d->targets = tt; 1539 d->ntargets = newn; 1540 1541 return &d->targets[oldn]; 1542 } 1543 1544 static struct aoetgt * 1545 addtgt(struct aoedev *d, char *addr, ulong nframes) 1546 { 1547 struct aoetgt *t, **tt, **te; 1548 1549 tt = d->targets; 1550 te = tt + d->ntargets; 1551 for (; tt < te && *tt; tt++) 1552 ; 1553 1554 if (tt == te) { 1555 tt = grow_targets(d); 1556 if (!tt) 1557 goto nomem; 1558 } 1559 t = kzalloc(sizeof(*t), GFP_ATOMIC); 1560 if (!t) 1561 goto nomem; 1562 t->nframes = nframes; 1563 t->d = d; 1564 memcpy(t->addr, addr, sizeof t->addr); 1565 t->ifp = t->ifs; 1566 aoecmd_wreset(t); 1567 t->maxout = t->nframes / 2; 1568 INIT_LIST_HEAD(&t->ffree); 1569 return *tt = t; 1570 1571 nomem: 1572 pr_info("aoe: cannot allocate memory to add target\n"); 1573 return NULL; 1574 } 1575 1576 static void 1577 setdbcnt(struct aoedev *d) 1578 { 1579 struct aoetgt **t, **e; 1580 int bcnt = 0; 1581 1582 t = d->targets; 1583 e = t + d->ntargets; 1584 for (; t < e && *t; t++) 1585 if (bcnt == 0 || bcnt > (*t)->minbcnt) 1586 bcnt = (*t)->minbcnt; 1587 if (bcnt != d->maxbcnt) { 1588 d->maxbcnt = bcnt; 1589 pr_info("aoe: e%ld.%d: setting %d byte data frames\n", 1590 d->aoemajor, d->aoeminor, bcnt); 1591 } 1592 } 1593 1594 static void 1595 setifbcnt(struct aoetgt *t, struct net_device *nd, int bcnt) 1596 { 1597 struct aoedev *d; 1598 struct aoeif *p, *e; 1599 int minbcnt; 1600 1601 d = t->d; 1602 minbcnt = bcnt; 1603 p = t->ifs; 1604 e = p + NAOEIFS; 1605 for (; p < e; p++) { 1606 if (p->nd == NULL) 1607 break; /* end of the valid interfaces */ 1608 if (p->nd == nd) { 1609 p->bcnt = bcnt; /* we're updating */ 1610 nd = NULL; 1611 } else if (minbcnt > p->bcnt) 1612 minbcnt = p->bcnt; /* find the min interface */ 1613 } 1614 if (nd) { 1615 if (p == e) { 1616 pr_err("aoe: device setifbcnt failure; too many interfaces.\n"); 1617 return; 1618 } 1619 dev_hold(nd); 1620 p->nd = nd; 1621 p->bcnt = bcnt; 1622 } 1623 t->minbcnt = minbcnt; 1624 setdbcnt(d); 1625 } 1626 1627 void 1628 aoecmd_cfg_rsp(struct sk_buff *skb) 1629 { 1630 struct aoedev *d; 1631 struct aoe_hdr *h; 1632 struct aoe_cfghdr *ch; 1633 struct aoetgt *t; 1634 ulong flags, aoemajor; 1635 struct sk_buff *sl; 1636 struct sk_buff_head queue; 1637 u16 n; 1638 1639 sl = NULL; 1640 h = (struct aoe_hdr *) skb_mac_header(skb); 1641 ch = (struct aoe_cfghdr *) (h+1); 1642 1643 /* 1644 * Enough people have their dip switches set backwards to 1645 * warrant a loud message for this special case. 1646 */ 1647 aoemajor = get_unaligned_be16(&h->major); 1648 if (aoemajor == 0xfff) { 1649 printk(KERN_ERR "aoe: Warning: shelf address is all ones. " 1650 "Check shelf dip switches.\n"); 1651 return; 1652 } 1653 if (aoemajor == 0xffff) { 1654 pr_info("aoe: e%ld.%d: broadcast shelf number invalid\n", 1655 aoemajor, (int) h->minor); 1656 return; 1657 } 1658 if (h->minor == 0xff) { 1659 pr_info("aoe: e%ld.%d: broadcast slot number invalid\n", 1660 aoemajor, (int) h->minor); 1661 return; 1662 } 1663 1664 n = be16_to_cpu(ch->bufcnt); 1665 if (n > aoe_maxout) /* keep it reasonable */ 1666 n = aoe_maxout; 1667 1668 d = aoedev_by_aoeaddr(aoemajor, h->minor, 1); 1669 if (d == NULL) { 1670 pr_info("aoe: device allocation failure\n"); 1671 return; 1672 } 1673 1674 spin_lock_irqsave(&d->lock, flags); 1675 1676 t = gettgt(d, h->src); 1677 if (t) { 1678 t->nframes = n; 1679 if (n < t->maxout) 1680 aoecmd_wreset(t); 1681 } else { 1682 t = addtgt(d, h->src, n); 1683 if (!t) 1684 goto bail; 1685 } 1686 n = skb->dev->mtu; 1687 n -= sizeof(struct aoe_hdr) + sizeof(struct aoe_atahdr); 1688 n /= 512; 1689 if (n > ch->scnt) 1690 n = ch->scnt; 1691 n = n ? n * 512 : DEFAULTBCNT; 1692 setifbcnt(t, skb->dev, n); 1693 1694 /* don't change users' perspective */ 1695 if (d->nopen == 0) { 1696 d->fw_ver = be16_to_cpu(ch->fwver); 1697 sl = aoecmd_ata_id(d); 1698 } 1699 bail: 1700 spin_unlock_irqrestore(&d->lock, flags); 1701 aoedev_put(d); 1702 if (sl) { 1703 __skb_queue_head_init(&queue); 1704 __skb_queue_tail(&queue, sl); 1705 aoenet_xmit(&queue); 1706 } 1707 } 1708 1709 void 1710 aoecmd_wreset(struct aoetgt *t) 1711 { 1712 t->maxout = 1; 1713 t->ssthresh = t->nframes / 2; 1714 t->next_cwnd = t->nframes; 1715 } 1716 1717 void 1718 aoecmd_cleanslate(struct aoedev *d) 1719 { 1720 struct aoetgt **t, **te; 1721 1722 d->rttavg = RTTAVG_INIT; 1723 d->rttdev = RTTDEV_INIT; 1724 d->maxbcnt = 0; 1725 1726 t = d->targets; 1727 te = t + d->ntargets; 1728 for (; t < te && *t; t++) 1729 aoecmd_wreset(*t); 1730 } 1731 1732 void 1733 aoe_failbuf(struct aoedev *d, struct buf *buf) 1734 { 1735 if (buf == NULL) 1736 return; 1737 buf->resid = 0; 1738 clear_bit(BIO_UPTODATE, &buf->bio->bi_flags); 1739 if (buf->nframesout == 0) 1740 aoe_end_buf(d, buf); 1741 } 1742 1743 void 1744 aoe_flush_iocq(void) 1745 { 1746 int i; 1747 1748 for (i = 0; i < ncpus; i++) { 1749 if (kts[i].active) 1750 aoe_flush_iocq_by_index(i); 1751 } 1752 } 1753 1754 void 1755 aoe_flush_iocq_by_index(int id) 1756 { 1757 struct frame *f; 1758 struct aoedev *d; 1759 LIST_HEAD(flist); 1760 struct list_head *pos; 1761 struct sk_buff *skb; 1762 ulong flags; 1763 1764 spin_lock_irqsave(&iocq[id].lock, flags); 1765 list_splice_init(&iocq[id].head, &flist); 1766 spin_unlock_irqrestore(&iocq[id].lock, flags); 1767 while (!list_empty(&flist)) { 1768 pos = flist.next; 1769 list_del(pos); 1770 f = list_entry(pos, struct frame, head); 1771 d = f->t->d; 1772 skb = f->r_skb; 1773 spin_lock_irqsave(&d->lock, flags); 1774 if (f->buf) { 1775 f->buf->nframesout--; 1776 aoe_failbuf(d, f->buf); 1777 } 1778 aoe_freetframe(f); 1779 spin_unlock_irqrestore(&d->lock, flags); 1780 dev_kfree_skb(skb); 1781 aoedev_put(d); 1782 } 1783 } 1784 1785 int __init 1786 aoecmd_init(void) 1787 { 1788 void *p; 1789 int i; 1790 int ret; 1791 1792 /* get_zeroed_page returns page with ref count 1 */ 1793 p = (void *) get_zeroed_page(GFP_KERNEL | __GFP_REPEAT); 1794 if (!p) 1795 return -ENOMEM; 1796 empty_page = virt_to_page(p); 1797 1798 ncpus = num_online_cpus(); 1799 1800 iocq = kcalloc(ncpus, sizeof(struct iocq_ktio), GFP_KERNEL); 1801 if (!iocq) 1802 return -ENOMEM; 1803 1804 kts = kcalloc(ncpus, sizeof(struct ktstate), GFP_KERNEL); 1805 if (!kts) { 1806 ret = -ENOMEM; 1807 goto kts_fail; 1808 } 1809 1810 ktiowq = kcalloc(ncpus, sizeof(wait_queue_head_t), GFP_KERNEL); 1811 if (!ktiowq) { 1812 ret = -ENOMEM; 1813 goto ktiowq_fail; 1814 } 1815 1816 mutex_init(&ktio_spawn_lock); 1817 1818 for (i = 0; i < ncpus; i++) { 1819 INIT_LIST_HEAD(&iocq[i].head); 1820 spin_lock_init(&iocq[i].lock); 1821 init_waitqueue_head(&ktiowq[i]); 1822 snprintf(kts[i].name, sizeof(kts[i].name), "aoe_ktio%d", i); 1823 kts[i].fn = ktio; 1824 kts[i].waitq = &ktiowq[i]; 1825 kts[i].lock = &iocq[i].lock; 1826 kts[i].id = i; 1827 kts[i].active = 0; 1828 } 1829 kts[0].active = 1; 1830 if (aoe_ktstart(&kts[0])) { 1831 ret = -ENOMEM; 1832 goto ktstart_fail; 1833 } 1834 return 0; 1835 1836 ktstart_fail: 1837 kfree(ktiowq); 1838 ktiowq_fail: 1839 kfree(kts); 1840 kts_fail: 1841 kfree(iocq); 1842 1843 return ret; 1844 } 1845 1846 void 1847 aoecmd_exit(void) 1848 { 1849 int i; 1850 1851 for (i = 0; i < ncpus; i++) 1852 if (kts[i].active) 1853 aoe_ktstop(&kts[i]); 1854 1855 aoe_flush_iocq(); 1856 1857 /* Free up the iocq and thread speicific configuration 1858 * allocated during startup. 1859 */ 1860 kfree(iocq); 1861 kfree(kts); 1862 kfree(ktiowq); 1863 1864 free_page((unsigned long) page_address(empty_page)); 1865 empty_page = NULL; 1866 } 1867